A NNALES DE L ’I. H. P., SECTION B
J. P ELLAUMAIL
A. W ERON
Integrals related to stationary processes and cylindrical martingales
Annales de l’I. H. P., section B, tome 15, n
o2 (1979), p. 127-146
<http://www.numdam.org/item?id=AIHPB_1979__15_2_127_0>
© Gauthier-Villars, 1979, tous droits réservés.
L’accès aux archives de la revue « Annales de l’I. H. P., section B » (http://www.elsevier.com/locate/anihpb) implique l’accord avec les condi- tions générales d’utilisation (http://www.numdam.org/conditions). Toute uti- lisation commerciale ou impression systématique est constitutive d’une infraction pénale. Toute copie ou impression de ce fichier doit conte- nir la présente mention de copyright.
Article numérisé dans le cadre du programme Numérisation de documents anciens mathématiques
http://www.numdam.org/
Integrals related to stationary processes and cylindrical martingales
J. PELLAUMAIL and A. WERON Ann. Inst. Henri Poincaré,
Vol. XV, n° 2, 1979, p. 127 -146.
Section B : Calcul des Probabilités et Statistique.
SUMMARY . The aim of this paper is to
analyse
arelationship
betweenintegrals
related tostationay
processes and squareintegrable martingales.
We consider the case when processes take their values in Banach spaces.
RESUME.
2014 L’objet
de ce travail est de mettre en evidence Fctroitc cor-relation existant entre diverses
integrales qui
interviennent dans1’etude,
d’une
part,
des processusstationnaires,
d’autrepart,
desmartingales cylindriques.
TABLE OF CONTENTS
0. Introduction ... 128
1. Quadratic integral ... 128
2. Non-negative measures ... 130
3. Operator valued functions and quadratic integral ... 134
4. Hilbert space case ... 136
5. Stochastic integral ... 138
6. Loynes spaces ... 140
7. Stationary processes ... 142
8. Square integrable cylindrical martingales ... 144
References ... 145
Annales de l’lnstitut Henri Poincare-Section B-Vol. XV, 0020-2347/1979/127i $ 4,00 0
127
INTRODUCTION
The aim of this paper is to
analyse
arelationship
betweenintegrals
related to
stationary
processes and squareintegrable martingales.
Weconsider the case when processes take their values in Banach spaces.
We establish here that
integrals (stochastic
anddeterministic)
used forthe above two classes of stochastic processes are
closely
related. Firstwe discuss such called
quadratic integral
for vector valued functions and for their tensorproducts.
After we direct our attension atnon-negative
measures with values in the space
(B’ @ 1 B’)’,
where B is a Banach space.In sections 3. 5 two
types
ofintegrals
foroperator
valued functions are described.Finally,
in the last sections 7 and 8 we show how the abovegeneral approach
is used forstationary
processes and for squareintegrable cylin-
drical
martingales.
In the paper we use the
langage
of tensorproducts
and the dilationtheory,
which has been
recently developed
fornon-negative
B-to-B’ valued ope- rators. Wehope
that thisgives
a newlight
forproblems
discussed here.The authors would like to express their
gratitude
to Professor M. Meti-vier for many valuable
suggestions
and conversations.1
QUADRATIC
INTEGRALIn this first
section,
we recall some definitions and results onintegration
with
respect
to vector measures and on tensorproduct:
weapply
thisnotions to the
study
ofquadratic integrals.
1.1. BILINEAR INTEGRAL.
We consider
- a set S and a
a-algebra L
of subsetsof S,
- a
complex
Banach space V and itstopological
dualV’,
- an additive V’-valued function M defined on
E,
- a V-valued function
f
defined on S.The first
problem
is to define theintegral fdM.
For an extensivestudy
on thisproblem,
we refer to[Bar ] .
~Annales de l’Institut Henri Poincaré - Section B
129 INTEGRALS RELATED TO STATIONARY PROCESSES
It is convenient for our purpose to consider the total variation which is defined as follows : if A
belongs
to Lwhere this supremum is taken over all the finite L-measurable
partitions
of A.
In
general situation,
this total variation is not finite(see 2.7),
but inthe
following
we often will suppose that this total variation vM is finite(or a-finite).
If the total variation vM is finite and
6-additive,
M isstrongly
a-additivedM and there exists a V’-valued weak
Radon-Nikodym
derivative M’ =-d
of M with
respect
to this R. N. derivative is V’-valued and for each element v ofV, ( M’, v ~
is the classical real R. N. derivativeof ( M, v ~
with
respect
to(briefly
w. r.t.)
vM.Conversely,
if this R. N. derivative M’exists w. r. t. a
positive
measure ,u, then the total variation vM is finite(see [Pel]).
In the
following,
when we say M is a measure, it means that M is a a-addi- tive measure.For the convenience we note the
following elementary
lemma :1.2. LEMMA.
Let M be a function defined on an
algebra
d and with values in the dual V’of a Banach space V. We suppose that the total variation ’!M of M is finite and
that,
for each element vof V,
the realfunction ~ v, M(.) )
is (7-additive.Then the total variation vM is a-additive.
The
proof
is left to the reader.1.3. TENSOR PRODUCT.
Let B and C be two
locally
convex vector spaces :sometimes,
we haveto use a bilinear continuous
mapping g
defined on B x C :often,
it isconvenient to consider this bilinear
mapping
as a linearmapping
definedon the « tensor
product »
B(x)
C.For definitions and
properties
of tensorproduct,
we refer to[Tre].
Let us recall
only
someproperties
when B and C are Banach spaces. Let D be a Banach space, BQ 1 C
is a Banach space suchthat,
for each D-valued bilinear continuousmapping g
defined on B xC,
there exists a D-valuedJ. PELLAUMAIL
linear
mapping f
defined on B@ 1 C
such that thefollowing diagram
iscommutative
where j
is a canonicalimbedding.
The norm in B@ 1
C is called theprojective
norm or the trace-norm.
There is an
isometry
from(B 01 C)
onto LN(B’, C),
the Banach space of linear nuclearoperators
from B into C with the trace norm; thisisometry
is the extension of the
mapping which,
to x@
y, associates theoperator
h -~ ~ x, h ~ . Y.
On the other
hand,
there is anisometry
from the dual(B 0 ~ 1 C)’
of(B 01 C)
ontoCL(B, C’),
the Banach space of continuous linearoperators
from B into C’ with the usual norm; thisisometry
associates to the linear formf
on B0~ 1 C,
the continuous linearoperator f
definedby
~ f (x)~ y ~ = f (x~ Y)~
1.4.
QUADRATIC
INTEGRAL.For
studying
second order processes(see [MaS ], [ChW 2]’
or[Wer3])
or
integral
w. r. t.quadratic
Doleans measures(see [MeP] ]
and[Met 1 D,
it is necessary to consider
integrals
of thetype XdMY* where M is a
a-additive function with values in the space of continuous antilinear
operators CL(B’, B")
and X and Y are functions defined on S and with values in B’.Actually,
M can be considered as a V’ =(B’ @ 1 B’)’-valued
additivefunction and
f
= X@
Y can be considered as a .B’@1 B’ =
V-valuedfunction; then,
theintegral XdMY*
= is aspecial
case ofthe
general integral
which was discussed in1.1; thus,
we can use all theclassical results on such bilinear
integrals.
In the next sections we will
study
moreprecisely
theintegral
X@ YdM,
when M is a
non-negative
measure."
2. NON-NEGATIVE MEASURE
In this section we shall consider the space V = B’
0i 1 B’
where B isa
complex
Banach space. If X: S -~ B’ and Y: S - B’ are measurablemappings
then ~ _ _ _ _ _ ~ _ w . __Annales de l’lnstitut Henri Poincaré - Section B
INTEGRALS RELATED TO STATIONARY PROCESSES 131
is measurable too. We are interested in measures with values in
(B’ 0i B’)’
= V’ which arenon-negative
i. e., VXeB’ ~0394~03A3Such measures in natural way arise in the
theory
ofstationary
processes(the
measure whichcorresponds
to the correlationfunction)
and in thetheory
of squareintegrable cylindrical martingales (quadratic
Doleans’measure),
see(§ 8)
below.We need the
following definition.
If B = H is a Hilbert space then non-negative
measureE( . )
with values in the set of allorthogonal projections
in H is called a
spectral
measure. ThusVA, Ai, A~
E E.2.1. LEMMA.
For each
non-negative
measure M with values in(B’ Q 1 B’)’
there exista Hilbert space
H,
anoperator
R ECL(B’, H)
and aspectral
measureE( . )
such that
Moreover if H is minimal i. e., H = then
H,
R andE( . )
. DEE
are
unique
up tounitary equivalence.
This means thatifM(A)
=where
E 1 ( . )
is aspectral
measure in a HilbertSpace H and .R1:
B’ -H 1,
then there exists an
unitary operator
U:H1
H such that R =UR1
1and
UE1 (Ll)
=E(A)U
for ‘d0 E E.Proof
- It follows from([W er 2], Prop. 3) by using
the fact that(B’ 01 1 B’)’
is isometric to CL
(B’, B").
The
triple (H, R, E( . ))
is called dilation of measureM( . ).
This dilation may beinterpretated
in thefollowing diagram:
.which shows that the measure
M( . )
is factorizedby
the HilbertSpace
H.2.2. LEMMA.
The Hilbert
Space
H in the abovediagram
isseparable
ifM(S)B’
is sepa- rable subset of B".J. PELLAUMAIL AND A. WERON
Proof
- SinceM(S)B’ = R*E(S)RB’
c R*M thenecessity
is clear.Conversely,
letM(S)B’
beseparable.
It is sufficient to prove that theimage R(C)
c H of the unit ball C in B’ isseparable.
be a sequence in C such is dense in
M(S)C.
We shall to prove
that ( Rbk}
is dense inR(C).
Let n ER(C). 3
b E C suchthat Rb = n.
Choosing
asubsequence (
such thatM(S)bkn M(S)b
in B" we have
which
completes
theproof.
In the lastequality
we use the fact thatE(S)
=IH
and
consequently M(S)
= R*R.2.3. REMARK.
Let vM be the total variation of M. Then
by
lemma 1.2 oo iff thereexists a non
negative
finite measure p on L suchthat ( X I (
for each X E B’.
Measures with such
property
was calledp-bounded
in[Mas ].
2.4. LEMMA.
The
following
conditions on measure M areequivalent:
i)
the total variation vM is infinite :over all finite collections of vectors from B’
with II
1 and allparti-
tions of S into a finite number of
disjoint
setsin E;
iii)
there exists a B’-valued measurable function X such thatSup II
1and
(X ~ X)dM = + 00.
s
Proof
- From the definition of the total variation of Mimmediately
follows that if vM = oo then
(ii)
holds. To prove the converse we observe that the normof II M(A)!! ( in (B’ 0i 1 B’)’
isequal
to the sup[M(A) ](X
@X),
where the supremum is taken over all vectors X from B’
with II
1.Thus
(ii) implies (i).
Is is easy to see from the definition of theintegral
f dM
that(ii)
isequivalent
to(iii).
Whichcompletes
theproof.
2.5. DEFINITION.
We say that M is
absolutely
continuous w. r. t. anon-negative
measure /1 on E if foreach x,
y EB’, [M( . ) ](x
@y)
« ,u.Annales de l’lnstitut Henri Poincaré - Section B
INTEGRALS RELATED TO STATIONARY PROCESSES 133
2.6. LEMMA
If for measure M the
image
of B’ underM(S) : M(S)B’
c B" isseparable,
then there exists a
non-negative
finite measure p on E such that M is absolu-tely
continuous w. r. t. J1.Proof - By
lemma 2 .1M( . )
has a minimal dilationM( . )
=R*E( . )R.
Moreover our
assumption
on Mimplies by
lemma 2.2 that the Hilbert space H in above dilation isseparable.
Observe that we have H = Let p be
given by
the formula :Ael
where
(en)
is CONS in H. Let us now use the fact thatE( . )
isspectral
measurein
CL(H, H).
Thus foreach n, ~ E(0394)en~2
isnon-negative
finite scalar-valued measure on E and
consequently j.1
isnon-negative
measure on(S, L)
and = 1.
Moreover, E( . )en 112
« j.1. Sincewe conclude that
Vh, g
E HBut
for x, y E B’, h = E( . )Rx
andg = E( . )Ry belong
to H and[M(.)](x @ y) = (R *E(. )Rx)(y) = (E2~ . E( )Ry) _ (E( . )h~ g)~
In the last
equalities,
we use the fact that forspectral
measureE2( . )
=E( . ).
Hence
by using
the above facts we obtain thatVx,
y E B’2.7. EXAMPLE.
Now we
give
anexample
of a measure M with values in(B’ @, B’)’,
where B’ = H is a
separable
Hilbert space, which isstrongly
a-additivebut its total variation v,~ is not a-finite.
Let S =
[0, 1 ]
x[0, 1 ],
~ =a-algebra
of the Borel sets of S, fi- theLebesgue
measure on S and a-the
Lebesgue
measure on[0, 1 ].
Weput
where % is the
a-algebra
of the Borel sets on[0, 1 ].
For each element A ~ 03A3 and for each finitefamily ( f , gi)i~l
of elements from H x H we defineNow, m
induces an additivemapping
M from E into(H Q 1 H)’.
IfA = B x
C,
whereB,
C c[0, 1 ],
andthen we have the
following inequality
for the norm in(H Q 1 H)’ :
which shows that the total variation of M is not a-finite.
It remains to prove that M is
strongly
a-additive. Let A be an element of ~. It is easy to see that the. norm ofM(A)
in(H Q 1 H)’
isequal
to thesupremum of where the supremum is taken over
all the
pairs ( f, g)
of elements from Hwith (
1 and( ~
1.But
by
theCauchy-Schwartz inequality,
we haveHence
and
consequently
M is a a-additive function on X onto(H 0 ~ 1 H)’
withthe
strong topology.
2.8. REMARK.
The above
example
shows that there exists a measure M : ~ --~(B’ Q 1 B’)’
which is
absolutely
continuous w. r. t. a a-finite measure, but its total variation vM is infinite.3. OPERATOR-VALUED FUNCTIONS AND
QUADRATIC
INTEGRAL3.1. NOTATIONS.
In this
section,
theassumptions
and notations are as inthe previous
section 2. More
precisely :
E is a
a-algebra
of subsets ofS,
B is a Banach space, M is a(B’
(x)1 B’)’-
valued measure defined on E with total variation vM, this total variation vM is finite and a-additive and M’ = -2014.
dvM
Moreover we suppose that B’ is
separable
and that G is a Banach space.Annales de l’lnstitut Henri Poincaré - Section B
135 INTEGRALS RELATED TO STATIONARY PROCESSES
3.2. ADJOINT OPERATOR
(see [Y os]
p.196).
Let x be an element of
L(B, G)
with domain D dense inB; then,
theadjoint operator
x* is an elementof L(G’, B’)
definedby
for
bED;
3 . 3 . WEAK * MEASURABILITY.
Let X be an
L(B, G)-valued
function defined onS;
we say that X isweakly*
measurableif,
for every element(b, g’)
of(B
xG’), ( X(b), g’ ~
is measurable. ’
Remark. The situation in this section is more
general
than in the pre- vious ones where we consider the case G = C andconsequently L(B, G)
= B’.For
studying
infinite-dimensional stochastic processes in Banach spaces thenecessity
arises ofusing
suchoperator
functions(see
sections 7 and8).
3.4. LEMMA.
Let u be an element of
(B’ 01 1 B’)’, g’
be an element of G’ and v be anelement of
L(B, G)
with domain D dense in B. Let v* be theadjoint
of v.Let
(Cn)n>
o be a sequence of elements ofB’,
dense in B’. LetBn
the vectorspace
generated by {
1, k, n. Let nn a be a linearidempotent
contraction from B’ ontoBn.
Then we have :Proof: By
ourassumptions,
~c" ECL(B’, BJ,
= C if C EBn
andif CeB’. Such
operator
existsaccording
to the Hahn-Banach theorem.
If we
consider v, g
and E >0,
there exist k > 0 and C EB~
such that~ v*(g) -
E ; thisimplies,
forevery n
k :Thus {03C0n v*(g)
convergesstrongly
tov*(g)
in B’.Now the lemma follows from the
continuity
of u and from thecontinuity
of the
mapping (x, y)
-~(x
(x)y)
for the « trace norm » on(B’ @ 1 B’).
3.5. PRELIMINARY PROPOSITION.
Let X be an
L(B, G)-valued
function defined on S andweakly*
measura-ble. For every element s of
S,
we suppose that the domain ofX(s)
is densein B. Let X* be the
adjoint
of X.Then,
for everyelement g
ofG’,
the func-tion
M’ { [~*(g) ]°’ ~
is a(real
nonnegative)
measurable function.Thus,
the
integral f M’ { [X*( g} ] ° 2 ~ . dvM
is well defined(finite
orinfinite).
Proof
- Let o be a sequence ofoperators
as in the above lemma.This lemma
implies :
But,
for everyinteger
n,M’ { [~n ~ X*( g) ] ° 2 ~
is measurableaccording
to the weak*
measurability
of X.Then,
the sameproperty
holds for~X*tg)]°2 ~
3.6. DEFINITIONS OF
~g
ANDJ~.
Let g be an element of G’.
We say that a
L(B, G)-valued
function X defined on S is an elementof
~~,
if :i)
X is weak*measurable,
ii)
for every element s ofS,
the domainof X(s)
is dense inB, iii) NR(X)
+ oo, whereMoreover,
we denoteby J~
the vector space of all elements X from such thatN~X)
+ oo, where4. HILBERT SPACE CASE
Here we sketch a construction of a
quadratic integral
which wasgiven
in
[Mas ]. H,
K areseparable
Hilbert space.Let X be an
L(H, K)
valued function onS ;
in[Mas ] X
is said measurable if there exists a sequence ofsimple
measurableCL(H, K)
fonctions suchthat V E S and for each x E H we have
lim !! X(S)x ))
= 0.n
Let M be a
non-negative CL(H, H)-valued
measure on E and M’ be aweak
Radon-Nikodym
derivative w. r, t. the total variation vM, whichby
the
assumption
is finite. We note that in[Mas ] the
author assumed that M isp-bounded.
Butby (2 . 3)
this isequivalent
to ourassumption.
Annales de l’Institut Henri Poincaré - Section B
137
INTEGRALS RELATED TO STATIONARY PROCESSES
4.1. DEFINITION
([Mas]).
The
pair (X, Y)
is Mintegrable
ifi) XM,1/2
and areCL(H, K)
valued and measurable(see above), ii) the
function(XM’1/2).(YM’1/2)*
isBochner vM integrable.
We denote
4. 2. DEFINITION.
~ ~
The functions
X,
Y with(X - Y)M’ 1 ~2
= 0 a. s.] are
identified.From the definition of
L2,M
one may see that4 . 4 . PROPOSITION
([Mas]).
L2,M
is a Hilbert space over thering CL(K, K)
with the normand moreover the
simple
functions are dense inL2,M.
4.5. PROPOSITION.
If H = K,
then the measure M has aspectral
dilation inL2,M’
Proof
- LetE 1( . )
be aspectral
measure inL2,M
defined as anoperator
of
multiplication by
the indicator1~.
Let R be the inclusion of K intoL2,M.
Then R* is an
orthogonal projection
onto K and for each A E L4.6. REMARK.
It is
interesting
to compare the aboveintegral
with one in[Met1],
Sect.
IV,
which is aspecial
case of theintegral
defined in(3 . 5) ;
we notethat
by (4 . 3)
the spaceL2,M
andK) (defined
in[Met1],
p.54)
are very similar.
Vol.
5. STOCHASTIC INTEGRAL 5.1. INTRODUCTION.
In this
section,
wegive
a construction of theintegral XdW in a general
context which includes stochastic
integrals
withrespect
to W where Wcan be a
stationary
process(see ]
and or acylindrical martingale (see [Met] ]
and(MeP D.
A similarapproach,
forstationnary
processes with values in Hilbert space
only,
can be found in[Mas ],
Inour
situation,
X is anL(B, G)-valued
function(or
process).
The assump- tions on W are as follows :5.2. NOTATIONS.
.. ’
Let H be a Hilbert space and B and G two Banach spaces where B’ is
separable.
Let d be analgebra
of subsets of the set S and E the6-algebra generated by
d. We denoteby
W an additive function defined on E and with values inCL(B’, H)
suchthat,
for every element(u, r)
of(B’
xB’),
and for every
pair (E, F)
of elements of d such that E n F =cjJ,
we have~ W(F)(v) ~H =
0.Then,
for every element(F,
u,v)
of(E
x B’ xB’)
we define :M is an additive
mapping
on d which induces an additivemapping
with values in
(B’ (X~ 1 B’)’
that we denote alsoby
M.We denote
by
vM the total variation of M(for
the norm in(B’ Q 1 B’)’).
We suppose that this total variation is finite and a-additive.
In this case, M can be extended into a
strongly
a-additivemapping,
defined on L and with values in
(B’ 01 1 B’)’
that we also denoteby
M.Then,
all theassumptions given
on M in section 3 are fulfilled. Asbefore,
we denote
by
M’ the R. N. derivative of M w. r. t. vM.5.3. 5i-SIMPLE FUNCTIONS.
We denote
by ~
the set of the functions such that X= xi .
tellA~~)
where(Xi)ieI
is a finitefamily
of elements ofL(B, G)
with domain dense in B and(A(i))iEI
is an associatedfamily
of elements of d.Annales de l’lnstitut Henri Poincaré - Section B
INTEGRALS RELATED TO STATIONARY PROCESSES 139
In this case, for every
element g
ofG’,
we define :and
XdW is
an element ofL(G’, H)
that we call theintegral
of Xw. r. t. W.
Of course,
if,
for every element i ofI, xt belongs
toCL(G’, B’),
then( BJ XdW)
is an element ofCL(G’, H) ;
but( BJ
can be an elementof
CL(G’, H)
for elementsxt
which do notbelong
toCL(G’, B’).
Moreover,
for every element(E, g)
of xG’),
we define : ,Then,
themapping
E(EXdW) is
an additivemapping
definedon A and with values in
L(G’, H).
The
problem
is to extend themapping
X XdW defined on ~ toa
sufficiently large
class of functions.For this extension the
following elementary
lemma is needed :S . 4 . LEMMA.
For every element X of ~ and for every element g of
G’,
we have :Proof.
2014 Xbelonging
to~,
we have X= 03A3xi.1A(i).
We can supposetel
that the sets are
pairwise disjoint ;
in this case, we have :by using
the fact thatW( . )
isorthogonal
ondisjoint
sets.Vol. n° 2-1979.
WERON
Then,
the definition of Mimplies
5.5. CONSTRUCTION OF THE INTEGRAL.
The lemma 5-4 above shows
exactly that,
for every element g ofG’,
the H-valued linearmapping
XH
XdW)(g)
defined on ~ is continuous for the semi-normNg
on 6 c~g
and thestrong topology
of H.Thus,
thismapping
can be extended into anunique
H-valued linearmapping
defined on the closure of lff in~g
for the semi-normNg.
Conse-quently
theintegral XdW is defined,
in a unique
way, for all the elements
of the closure ~203C0
of 6 in 2;,
i. e. this integral
is defined for all the L(B, G)-
valued functions X which are
weakly* measurable, X(s)
has a domaindense in B for every element s of S and such that
N1[(X)
+ 00(see
3-6above). Moreover,
if Xbelongs
to6i, XdW is an element of CL(G’, H).
6. LOYNES SPACES
Our attention is now directed at spaces on which a vector-valued inner
product
can be defined.Let
Z
be acomplex
Hausdorfftopological
vector spacesatisfying
thefollowing
conditions :(6 .1 ) Z
has an involution : i. e. amapping
Z --~Z +
ofZ
to itself with theproperties
(6 . 2)
there is a closed convex coneP
inZ
such thatP
n -P
= 0 ; thenwe define a
partial ordering
inZ by writing Z,
ifZl - Z,
EP ;
Annales de l’lnstitut Henri Poincaré - Section B
INTEGRALS RELATED TO STATIONARY PROCESSES 141
(6 . 3)
thetopology
inZ
iscompatible
with thepartial ordenng,
in thesense that there exists a basic
set { No }
of convexneighbourhoods
of theorigin
such that if x ENo
andx a y a
0 then y ENo ;
(6 . 4)
ifxeP
then x =x+ ;
(6 . 5) Z
iscomplete
as alocally
convex space ;(6 . 6)
if... > 0
then the sequence xn isconvergent.
We remark that it is clear that these conditions are satisfied
by
thecomplex
numbers and
by
the space afall q
x q matrices with the usualtopologies.
Now,
suppose thatH
is acomplex
linear space. A vector innerproduct
on
H
is a map x, y -[x, y] ]
fromH
xH
into an admissible spaceZ (i.
e.,satisfying (6.1)-(6.6))
with thefollowing properties :
for
complex a
and b. When a vector innerproduct
is defined on the spaceH
there is a natural way in whichH
may be made into alocally
convextopo- logical
vector space.Namely
a basic set ofneighbourhoods
of theorigin { uo }
is definedby
(6.11)
DEFINITION.The space
H
which iscomplete
in thetopology
definedby (6.10)
withthe
admissible
spaceZ satisfying
conditions(6 .1 )-(6 . 6)
will be called aLoynes
space. This space was introduced in[Loy].
A
complex
Hilbert space and the space of all p x q matrices with the usualtopologies
aresimple examples
ofLoynes
spaces.(6.12)
PROPOSITION([ChWJ).
Let B be a
complex
Hilbert space and H acomplex
Hilbert space. Then the spaceCL(B, H)
with thestrong operator topology
is aLoynes
space if we defineA +
as anoperator adjoint
toA, P
as the set of allnon-negative operators
and[A,C]=C+A
with values in an admissible space
CL(B, B)
with the weakoperator topo-
logy.
Using
the remark from sect.1,
we have thatCL(B, B)
is isometric to(B’ 0i 1 B’)’.
We will see below that the aboveLoynes
spaceL(B, H)
isclosely
related tostationary
processes if H = and to2-integrable cylindrical martingales
if H = M-the Hilbert space of real square inte-grable martingales.
7. STATIONARY PROCESSES
Let H =
L~(Q). By
second order stochastic processes with values ina Banach space
B, briefly B-process,
we mean thetrajectory Xt
in theLoynes
spacesL(B’, H).
We refer the reader to[Wer1] for
the motivationand
examples
of such processes. Its correlation functiontakes values in the admissible space
(B’ §) ~ 1 B’)’.
If T is an Abelian group then a
B-process
isstationary
if the functionof two variables
K(t, s) depends only
on(t
-s).
In the case when T is atopological
group, we will assume thatK(t)
isweakly
continuous i. e.,~b~B’,
K(t) (b @ b)
is continuous. If T is alocally compact
Abelian(LCA) group
then wehave,
cf.]
where
E( . )
is aspectral
measure in the spaceHx
=Sp { X tb, t
ET,
b EB’ ~ ,
defined on the Borel
a-algebra
of the dual group Gand ~ t, g ~
-a valueof the
character g
E G atpoint t.
In the case T = R we have G = R and( t, g ~ - eitg, t, g
ER, similary
for T =Z,
G =[0, 2n] ] and ~ t, g ~ - eitg,
[0, 2n ].
We note that the
integral (7 .1 )
is aspecial
case ofintegral
w. r. t. anorthogonal
measure, which was studied in section 5. Indeedis an
orthogonal
measure with values inCL(B’, H).
Here X isequal to ~ t,
g~.
(7.2)
PROPOSITION.B-process Xt
isstationary
if its correlation function has the formAnnales de l’lnstitut Henri Poincaré - Section B
INTEGRALS RELATED TO STATIONARY PROCESSES 143
where M is
non-negative (B’ @ ~ B’)’-valued
measure defined on the Borela-algebra BG
of the dual group G.Proof
- IfXt
isstationary
then from the definition of the correlation function and(7 .1 )
we haveThus if we
put
we obtain
(7 . 3). M( . )
isnon-negative because, according
to theproperties
of the
spectral
measureE(s) (cf.
Sect.2),
we have thatConversely,
LetM( . )
be anon-negative (B’ @ 1 B’)’-valued
measure onBG.
Then, by
lemme 2.1 there exist aHilbert-space H,
anoperator
R ECL(B’, H)
and a
spectral
measureE 1 ( . )
in H such thatWe may take H = Let we
put Ut = G t, g > E 1 (dg)
for each t E T.
Then, Ut
is aunitary operator.
Now we define a processXt by
thefollowing
formula
Consequently, Xt
has the form(7.1)
and isstationary.
(7.4)
REMARK.Observe that
W(.)
=E(. )Xe
is anorthogonal
random measure andthat each
stationary B-process
has therepresentation
asintegral (sto- chastic)
w. r. t. anorthogonal
random measureConsequently,
the measure M connected with the correlation function of astationary B-process (by 7 . 2)
has the form :Vol. 2-1979.
We note that the above fact is
general
and holds for allnon-negative (B’ @ 1 B’)’
valued measure on(S, ~).
It followsimmediately
from thedilation theorem
(Lemma 2.1).
(7.5)
PROPOSITION.If
M( . )
is anon-negative (B’ 01 1 B’)’-valued
measure on(S, E)
then thereexists a Hilbert space H and a
orthogonaly CL(B’, H)-valued
measureW( . )
on(S, L)
such that for each A e EMoreover,
if H is minimal H =V : W(A)B’}
then H and W areAeX
determined up to
unitary equivalence.
(7.6)
EXAMPLE.Let B be a Hilbert space. Then stochastic
integral (as
in7.4)
is definedas
mapping
fromL2,M
intoH° :
closedsubspace
of Hilbert-Schmidt ope- ratorsspanned by Ai
ECL(H), ti
E T and N isinteger
cf.[MaS ].
We note
only
that in view of(2.1)
and(4. 5)
this stochasticintegral
realizedunitary equivalence
between twospectral
dilations of the measure M definedby
the correlation function of thestationary
process.8. CYLINDRICAL MARTINGALES
Let
(Q, ff, P,
be aprobabilized
stochastic basis(see,
forexample [MeP2 ]).
Let be the space of all thecomplex
squareintegrable
martin-gales (with respect
to this stochasticbasis) ;
if(Ut)teT
and(Vt)teT
are two suchmartingales,
we can defineFor this calar
product ( .,. ) ,
~~r is an Hilbert space.Let B be a Banach space.
Following Prop. (6.12), CL(B’, ~~~)
is aLoynes
space. Then, an element
W
ofCL(B’, ~~)
is called asquare-integrable cylindrical martingales (see [Met] ]
or[MeP1 D.
We define S = Q x T and we denote
by Z
thea-algebra
of thepredictable
Annales de l’lnstitut Henri Poincaré - Section B
145 INTEGRALS RELATED TO STATIONARY PROCESSES
sets
(with respect
to the stochastic basisgiven above).
We denoteby d
the
algebra,
included inE, generated by
the sets(D
x]r, s])
withreT,
s E T and For such an element F = D x
]r, s] ]
and for every element u ofB’,
we denoteby W(F)(u)
the realmartingale
which is definedby:
Then,
W is an additive function which satisfies all theproperties given
in 5.2 above with H = -e7.
In this case, the function
M,
associated to W as in 5-2above,
is the qua- dratic Doleans measure of thecylindrical martingale
W. If X is aweakly*
predictable
process(with
values inL(B, G)),
X can be considered as aweakly*
E-measurable
function;
in this case, XdW is the stochasticintegral
of Xwith
respect
to W."
For other details and
results,
see[Met ] and [MeP 1 ].
REFERENCES
[Bar] R. G. BARTLE, A general bilinear vector integral. Studia Math., t. 15, 1956, p. 337-352.
[ChW1] S. A. CHOBANYAN and A. WERON, Stationary processes in pseudo-Hilbertian
space (in Russian), Bull. Acad. Polon., t. 21, 1973, p. 847-854.
[ChW2] S. A. CHOBANYAN and A. WERON, Banach space valued stationary processes
and their linear prediction. Dissertationes Math., t. 125, 1975, p. 1-45.
[Dou] R. G. DOUGLAS, On factoring positive operator functions. J. Math. Mech.,
t. 16, 1966, p. 119-126.
[DuS] N. DUNFORD and J. T. SCHWARTZ, Linear operators. Part. 1. Intersciences Publishers, 1967.
[Loy] R. M. LOYNES, Linear operator in Vh-spaces, Trans. Amer. Math. Soc., t. 116, 1965, p. 167-180.
[Mas] V. MANDREKAR and H. SALEHI, The square integrability of operator-valued
functions with respect to a non-negative operator valued measure and the Kolmogorov isomorphism theorem. Indiana University Mathematical Journal.
vol. 20, n° 6, December 1970.
[Mas] P. MASANI, Quasi isometric measures and their applications. Bull. Amer. Math.
Soc., vol. 76, n° 3, 1970.
[Met1] M. METIVIER, Integration with respect to processes of linear fonctionnals. Séminaire
de Probabilité de Rennes, 1975.
[Met2]
M. METIVIER, Cylindrical martingales and stochastic integrals. Lecture on the conference. Probability Theory on vector spaces. Trzebieszowice, 1977 (non published).[MeP1] M. METIVIER and J. PELLAUMAIL, Cylindrical stochastic integral. Séminaire de Rennes, 1976.
[MeP2] M. METIVIER and J. PELLAUMAIL, A basic course on stochastic integration.
Séminaire de Rennes, 1978.
2-1979.
[Pel] J. PELLAUMAIL, Sur la dérivation d’une mesure vectorielle. Bull. Soc. Math.
France, t. 98, 1970, p. 305-318.
[Tre] F. Treves, Topological vector spaces, distributions and kernels. Academic Press, 1967.
[Wel] J. N. WELSH, Constructions of the Hilbert space L2,M for operator measures.
Proc. Symp. on vector and Operator measures. A. Utah, 1972.
[Wer1] A. WERON, Prediction theory in Banach spaces. Lecture Notes in Math., vol. 472, 1975, p. 207-227.
[Wer2] A. WERON, Remarks on positive-definite operator valued functions in Banach spaces. Bull. Acad. Polon., t. 24, 1976, p. 873-876.
[Wer3] A. WERON, Second order stochastic processes and the dilation theory in Banach
spaces (to appear).
[Yos] K. YOSIDA, Functionnal analysis. Second printing, Springer Verlag, 1966.
(Manuscrit reçu le 5 mars 1979)
Annales de l’Institut Henri Poincaré - Section B